EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Revert "perf: optimize qwen-vl with symm mem allreduce" (#11436)

Browse Source
This commit is contained in:
Cheng Wan
2025-10-10 12:30:05 -07:00
committed by GitHub
parent af96ca1136
commit 52fcbbb8bd
5 changed files with 17 additions and 82 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/managers/schedule_batch.py
View File

@@ -1766,11 +1766,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
self.seq_lens_cpu = self.seq_lens_cpu[keep_indices]
self.orig_seq_lens = self.orig_seq_lens[keep_indices_device]
self.out_cache_loc = None
if isinstance(self.seq_lens_cpu, torch.Tensor):
# CPU tensor
self.seq_lens_sum = int(self.seq_lens_cpu.sum().item())
else:
self.seq_lens_sum = int(np.asarray(self.seq_lens_cpu).sum())
self.seq_lens_sum = self.seq_lens.sum().item()
self.output_ids = self.output_ids[keep_indices_device]
self.return_logprob = any(req.return_logprob for req in self.reqs)
if self.return_logprob: